How to Remove Vocals from a Song | WaveCraft

The short answer

Load a track into WaveCraft, run vocal separation, and you get two files: vocals and instrumental. The instrumental is the karaoke version; the vocals file is a solo you can study, remix or record over. The pass runs on your device — the audio never leaves the phone, and it works with no connection at all.

What that pass does to the mix matters more than the button that starts it. WaveCraft is free to record and edit with; unlimited separation and lossless export are Pro. WaveCraft is on the App Store.

Why you are searching for this

Usually one of three situations: you want a backing track to sing over and do not own the recording; you cannot hear a part under the rest of the mix; or you got a hollow, echoey result from another tool and cannot tell whether the tool is broken or the technique is.

The third decides whether you should keep trying. That hollow, bassless sound is not a settings problem. Cheap vocal removers assume the singer sits dead centre and subtract one stereo channel from the other — which works on old hard-centre recordings and destroys everything else living in the middle. Separation that reconstructs cancels nothing, so it has nothing to hollow out.

What the pass actually does to the mix

Not magic, and not subtraction. The audio is converted to a spectrogram — the same short-time frequency picture a spectrum analyser draws — on a 4,096-point transform with a 1,024-sample hop at 44.1 kHz. Sources at another rate are normalised to it, because that geometry and sample rate are what the model was trained on; feed it something else and you are asking it to generalise rather than recognise.

An on-device model reads that picture and estimates, for every moment and every frequency, how much belongs to the singer. Two gain sets come out: one for the vocal, one its complement. The vocal track is the mix multiplied by the first; the instrumental is the mix multiplied by the second. Three consequences follow:

• Nothing is cancelled. The instrumental is not "the original minus the singer" — it is a rebuilt signal. Whatever the model misjudges therefore survives in both files at once, which is why bleed shows up on each side rather than being deleted from one.
• The two outputs are complementary by definition. A gain and its complement sum to one at every frame and every bin, so vocals plus instrumental reconstructs the original mix sample for sample. That property is worth more than it sounds — see the four questions below.

Analysis runs in blocks with a small context overlap, and cancelling mid-pass deletes the half-written files rather than leaving a truncated one behind.

Judging your source before you press the button

Separation is estimation, so the source decides the outcome more than any control does.

Separates well

• A studio mix where the lead vocal sits clearly above the instruments.
• Modern dense pop with a competently mastered vocal.
• Anything where vocal and instruments occupy different frequency territory at the same moment.

Separates acceptably, with audible bleed

• Reverb-heavy vocals. The reverb tail is vocal too, and some of it lands in the instrumental — most audible on sustained notes and after the last word.
• Older or lo-fi recordings mixed with the singer hard-panned.

Separates badly

• Heavily distorted or screamed vocals, where the processing that makes the vocal loud also makes it spectrally identical to the instruments.
• Extremely compressed masters, where vocal and instruments share the same frequencies.
• Dense arrangements with several singers overlapping. The model has to pick one, and it will.

None of these are settings problems, and none improve by running the pass twice.

The four questions everyone asks afterwards

Does it damage the rest of the song? With a reconstruction-based separator, no — there is no subtraction step that can eat your bass. What you can lose is fidelity: material attributed to the wrong source is redistributed across both outputs. If the instrumental sounds thin, the fix is the mix, not the separator.

Do the two files add back up? Yes, sample for sample. Practically this is the most useful thing to know, because it makes separation reversible rather than lossy. Put the instrumental under a video, a live band or your own take, and mute and unmute to A/B against the original instantly. If the stack comes out permanently thinner than the source, that is a defect in the tool, not in your project.

Is the separated vocal cleaner than what was recorded? No — and this is where people are most disappointed. It is a reconstruction of the same performance: same timing, same takes, same breaths. At best it removes what was in the way. For a genuinely cleaner vocal, denoise and EQ the take you recorded instead — see how to denoise a recording on iPhone.

Can I get a different take, or a vocal that was never sung? No. Separation cannot invent a harmony, fix a pitch or replace a singer. If you want that, you overdub — and the separated vocal is the reference you would practise against.

Doing it on a phone, start to finish

1. Open WaveCraft and start a new project.
2. Import the audio file. If the source is a video, import the video directly — WaveCraft extracts the audio track first, which gives you an audio timeline instead of a video one.
3. Select the track and run vocal separation from the AI Tools tab, alongside denoise, vocal enhance, loudness normalization and auto-mix.
4. Wait for the pass. Two tracks appear in your timeline.
5. Mute the one you do not want.

Then treat them like any other material in the project: unlimited tracks, split and trim, fades, speed change, and mixing with 20+ effects plus mastering presets. That step is not decoration — the instrumental is a stem, not a finished track. It arrives at the source mix's level carrying the source mix's balance and compression, so giving it its own EQ and level is what turns "the song minus the singer" into a usable backing. Export as WAV, AAC, ALAC or FLAC, up to 32-bit float; go lossless if the stems go into a DAW later.

What the two files are actually good for

Sing or play over the instrumental, with a short fade at the head and tail where residual vocal is most audible. Before publishing anything built this way, read the licensing section of our karaoke guide — removing the vocal does not change the copyright status of the recording. Slow a separated vocal down to learn a part, or layer your own take over it for a cover. If you want the drums on their own, that is a larger job than a two-way split — see vocal separation vs stem separation.

What WaveCraft does differently

• It runs on the device. No upload, no queue, no waiting on a server. It works on a plane, and your audio never leaves the phone. There is no account and no sign-in at all.
• The separation is complementary, not lossy. Vocals plus instrumental sums back to the original sample for sample, so every experiment with the instrumental is reversible. Plenty of separators quietly discard energy during reconstruction and hand you a mix that is permanently thinner than the source.
• The stems land in a real multitrack project. They are tracks on a timeline, not two exported files you have to re-import into a desktop DAW, so cleaning bleed, levelling and mastering the result happen in the same session — where denoise, vocal enhance, loudness normalization and auto-mix already sit next to separation, along with video audio extraction and export in WAV, AAC, ALAC or FLAC up to 32-bit float.

And the honest limits:

• Separation is estimation. Bleed survives, most audibly on sustained notes and reverb tails.
• It cannot improve a performance or produce a take that was never recorded.
• Pro is what makes this a workflow rather than a demo: $4.99/month, $24.99/year, or a one-time $49.99 Lifetime purchase with Family Sharing. Pro unlocks unlimited separation, denoise and auto-mix, all 20+ effects, every mastering preset, multi-track mixing, lossless WAV/ALAC/FLAC export, and no ads. A 7-day free trial is offered on the monthly and yearly plans for new subscribers; if you have already used it, the button reads Subscribe with no trial, so there is no surprise charge. The free tier lets you record and use the core editor first.

The decision you can make now

Three questions, in order:

1. Do you want a backing track, a reference vocal, or re-mixable stems? The first two come out of one pass. Stems cost more and are worth it only if you intend to re-mix.
2. Does your source have a voice sitting clearly on top? If yes, run it and expect a good result. If it is crushed, screaming, or several singers at once, expect bleed.
3. Do you need this with no signal, and without handing a copyrighted recording to a server? If either matters, on-device processing is the entire answer.

WaveCraft runs the split on your iPhone — get it on the App Store. Listen on headphones before you judge: the result is good enough to practise against and sing over, not to fool anyone about who was singing.